Support Vector Machine For Fault Diagnosis

Statistical learning theory (SLT) is based on the structural risk minimization (SRM) principle, and it is a new set of theory system, which specially aims at machine learning issues under the circumstances of small-sample. Based on this SLT, supporting vector machine (SVM) method has been developed as a new machine learning algorithm and also practical applications of SLT. At present, it is mainly used in the fields of pattern recognition, regression and probability density function estimation, etc. Fault diagnosis is part of pattern recognition issues, in this essay, SVM is applied to establish two-fault classifier and multi-fault classifier of materials, and the multi-variety algorithm of SVM is also discussed. Examples are used to make a study of the validity of SVM in fault diagnosis, and to make a comparison with classifier of neural network (NN).In this essay, some important concepts about SLT are firstly introduced i.e. SRM and VC dimension to illustrate that SVM holds excellent capability of generalization. In addition, the concept and process of answering of SVM is also introduced to illustrate that SVM is a sort of convex optimization issue whose answer has characteristics of global optimum. SVM method is presented on the condition of linear classifier, and has been developed as an effective way in solving problems of nonlinear pattern recognition. In consideration that the problem of classification existing in reality is always nonlinear, and unable to be separated completely, a minute explanation of C -support vector classification ( C - SVC ) is given andν-support vector classification (ν- SVC ) --a method for improvement in which parameterνhas objective significance is also introduced to avoid the disadvantage that parameter C in C - SVC is of no exact significance. Finally, some fault diagnosis cases about rolling contact bearing, oil pumping unit, especially twin-screw extruder are provided to demonstrate the establishment of multi-fault classifier. The property between C- SVC andν- SVC and two kinds of agorthm between"one-against-one"and"one against the other"are made a comparison respectively and the simulation results are given. Meanwhile, the differences of generalization capability between SVM and NN on the condition of small-sample are reflected when NN is applied to the same samples.The fault diagnosis cases are indicated that the two sorts of multi-variety algorithm of SVM have quite different diagnosis results in the same samples. Besides, with application of small-sample fault diagnosis, SVM is more adaptive and holds better generalization capability compared with NN.